Research On Efficient Deep Learning Based Automatic Modulation Recognition For Radio Signal

With the development of radio communication technology,spectrum resources are becoming more and more scarce in military,commercial,civil and other fields.The modulation recognition technology of cognitive radio signal has become an important technology that can effectively assist the scientific reconfiguration of wireless system and alleviate the shortage of spectrum.In recent years,deep learning has made great progress in the automatic modulation classification task of cognitive radio.However,in non-cooperative communication channel,the accurate identification of modulation types under low SNR is still a research difficulty and hotspot in this field.Complex interference(oscillator drift,clock drift,noise)in real communication channel will significantly reduce the performance of existing classification networks for radio signal recognition.In order to overcome the problems of recognition under low SNR and complex interference mentioned above,this study is based on efficient deep learning,conducted from learning efficiency and inference efficiency.Firstly,an architecture is proposed based on hybrid attention mechanism for automatic modulation classification,named RCN(Radio-Channel-Net).Two types of pre-modules are brought to realize the synchronization and standardization of raw signals and explore more features between channels,so as to obtain a transformed multichannel version of original signals with“cleaner and stronger” features,and then flow into the final network to get modulation results.The experimental results show that when the SNR is higher than-7d B,the RCN achieves the best performance in all the compared models;93.448% at 0d B,2.7% higher than CLDNN;97.560% at 20 d B,8.2% higher than Res Net.Based on the RCN model,two visualization technologies are used to guide the optimization research of resource consumption savings.Two strategies are compared in reducing model memory resources: a model compression method of channel pruning based on different modulation categories and a data compression method of wrapped feature selection based on network and autoencoder.Finally,the model compression strategy in the study gives 2 × reduction in model size with 1% accuracy loss.The feature selection strategy based optimization suffers only 0.5% accuracy loss with 79%parameter scale of the same original model.And compared with the former strategy,the model size is 10 times smaller and the overall accuracy is about 3% lower.